Bomb threats against an airline

• In 1983, a caller made a series of bomb threats by telephone to Pan American Airlines, and some calls were recorded
• The police charged a suspect, Paul Prinzivalli
• His defence sent the threat calls, and a recording of Prinzivalli, to the sociolinguist William Labov

Two vowels

• The caller had a Boston accent: he used almost the same vowel in bomb and off
• Prinzivalli had a New York accent, which keeps the two vowels apart
• Prinzivalli was acquitted; the judge named the phonetic evidence as crucial for reasonable doubt

Two vowels

Simulated values, for illustration: the caller merges the two vowels, the suspect keeps them apart.

What the case shows

• A questioned sample (who is speaking?) is compared with a known sample (a speaker whose identity we know)
• Speaker comparison is the most common type of case in forensic phonetics (session 4)
• Experts use auditory, acoustic and sometimes automatic methods
• They analyse speech sounds, like the vowels here, and features of the voice such as intonation and voice quality

Today

A refresher of the phonetics behind the rest of the course, with a forensic focus:

1. How speech is produced: the source-filter model, segments and prosody
2. How speech is measured: waveforms, spectrograms, formants, f0, voice quality and tempo
3. Which features forensic phoneticians actually analyse

The source-filter model

• Air pushed up from the lungs sets the vocal folds vibrating at the glottis. This vibration is the source
• The oral and nasal tracts then shape the sound, acting as a filter. Moving the tongue, lips and velum changes their resonances
• The filter works a bit like a stereo equaliser, boosting some frequencies and damping others

Source, filter, output

Simulated spectra: the filter’s peaks are the formants.

Notation

Airstreams

Most speech uses an egressive pulmonic airstream, with air pushed out from the lungs. Some sounds use other airstreams: clicks, for example, draw air inwards. In English, clicks are not speech sounds (tut tut, or spurring on a horse), but in some African languages, such as Xhosa, they are consonants.

Segments and prosody

Segments are individual consonants and vowels. Prosody, or suprasegmentals, covers features that stretch over more than one segment, such as stress, intonation, rhythm, tempo, tone and voice quality. Forensic phoneticians analyse both.

Consonants

Three parameters, often abbreviated VPM:

• Voicing: do the vocal folds vibrate? fuzzy (voiced /z/) vs fussy (voiceless /s/)
• Place: where is the constriction? From bilabial (lips) to glottal (vocal folds)
• Manner: how is the air obstructed?

Manners of articulation

• Obstruents: the airflow is blocked or turbulent
  • plosives (full closure, then release), fricatives (turbulent airflow), affricates (closure, then slow release with friction)
• Sonorants: air escapes easily
  • nasals (velum lowered), approximants (lateral /l/, central /ɹ/), trills, taps and flaps, and all vowels

Try it

Describe each consonant by voicing, place and manner:

• the first sound of mother
• the last sound of beach
• the sound written in sing
• the sound between the vowels in American English city

Vowels

Three parameters:

• Tongue height: close (high) to open (low)
• Tongue position: front, central, back
• Lip rounding: rounded or unrounded

Lexical sets

• Wells (1982) named groups of words after a keyword: PRICE stands for all words with the vowel of price (guide, might, nine)
• Accents differ in how they pronounce a set, but usually affect all its words in the same way
• Keywords work for every accent, unlike phoneme symbols, which imply the standard accent

Fundamental frequency and pitch

• f0 (“F zero”) is the number of vocal fold vibrations per second, in Hertz, so it is a measurement. Pitch is how high or low we hear a voice, which makes it a percept
• When pitch carries meaning over a phrase, we call it intonation. The dog chases the cat with a fall is a statement; with a rise, a question
• In tone languages, pitch also distinguishes words. Thai /kʰa/ with a high tone means “to trade”, with a falling tone “to kill”, with a rising tone “leg”

Typical f0

Simulated speakers. Typical values: men 100–125 Hz, women 180–220 Hz, children above 300 Hz. Longer, heavier vocal folds vibrate more slowly.

Voice quality

The shape and tension of the glottis change the timbre of the voice:

• Modal voice: regular vibration
• Breathy voice: the glottis does not close fully, air escapes
• Whisper: no vibration; turbulent air through a narrow opening
• Creaky voice: slow, irregular vibration of the edges of the folds

Prominence and rhythm

• Prominence: stressed syllables stand out, usually by being louder, longer and higher. IMport (noun) vs imPORT (verb)
• Languages are traditionally classed as stress-timed (stressed syllables at roughly regular intervals, as in English) or syllable-timed (syllables of roughly equal length, as in French). In reality rhythm is a continuum, and no language fits either type perfectly

Rhythm

Simulated syllable durations. English shortens unstressed syllables so that stressed ones come at regular intervals.

Speech tempo

• Articulation rate: syllables per second, without pauses
• Speaking rate: syllables per second, with pauses
• Fast speakers of English produce about 7–8 syllables per second, slow ones about 3–4
• Rates depend on the language’s syllable structure: mostly CV syllables (Japanese) give higher rates than many consonant clusters (Polish)

Speech tempo

28 syllables in 6.2 s, of which 4.79 s are speech:

• Articulation rate: 28 / 4.79 = 5.8 syllables per second
• Speaking rate: 28 / 6.2 = 4.5 syllables per second

Praat

• Free software for speech analysis by Paul Boersma and David Weenink (University of Amsterdam), and the standard tool for students, researchers and forensic experts. Praat is Dutch for “talk”
• Measures f0, formants, intensity and voice quality; annotates recordings; draws figures

The waveform

Simulated signal. Voiceless sounds are aperiodic; voiced sounds repeat with every vibration of the vocal folds.

From waveform to frequencies

Fourier: every complex periodic wave is a sum of sine waves at whole multiples of f0, the harmonics.

The spectrogram

• Time on the x-axis, frequency on the y-axis, intensity as darkness
• Each vertical striation is one opening and closing of the glottis (a glottal pulse)
• The vocal tract keeps changing shape, so the pattern of energy keeps changing: a running record of the filter

Consonants in the spectrogram

Resonance and formants

• Tap a wine glass: its sound depends on its shape and how much air it holds. The vocal tract works the same way
• Formants: the resonances of the vocal tract, dark horizontal bands in a spectrogram
• F1 goes down as the vowel gets closer: close [i] low F1, open [a] high F1
• F2 goes up as the vowel gets fronter: front [i] high F2, back [u] low F2
• F3 reflects lip rounding; F4 and above often carry speaker-specific information

Bigger tract, lower formants

Approximate cardinal vowel formants for an adult man (Catford, 1988), scaled up for illustration. The vowel space keeps its shape but moves up in frequency.

Measuring f0

• Praat’s voice report gives the average f0 (mean, median, mode), its standard deviation and its range for a selected stretch of speech
• But every f0 value is an estimate: the algorithm can be wrong, especially with noise, unusual voice qualities or poor recordings

When the tracker gets it wrong

Simulated f0 track of an adult man. Mean with errors: 114 Hz; without: 109 Hz. Median with errors: 109 Hz.

Check the numbers

• An adult man speaking normally should give f0 values around 100–130 Hz
• Values around 200 Hz could be real (a very high register, or falsetto), or the tracker doubling the true value
• Halved values work the other way
• Always look at the track, and listen, before trusting a summary statistic

Discussion

In a bomb threat call, the caller’s mean f0 comes out at 190 Hz. The suspect, interviewed calmly at the police station, has a mean of 110 Hz. The defence argues that the caller must be someone else.

• What could explain the gap without a second speaker?
• What would you look at, and listen to, before taking either side?

Measuring voice quality

• Jitter measures how much the duration of one vibration cycle varies from the next. In healthy young adults it is about 0.5–1%
• Shimmer is the same idea for amplitude
• The harmonics-to-noise ratio (HNR), in dB, falls as a voice gets noisier, hoarser or breathier. Values under 7 dB usually indicate a vocal pathology
• H1–H2, the difference between the first two harmonics, is used as a measure of breathiness and creakiness

Jitter

Simulated cycle durations of a voice at 125 Hz (8 ms per cycle).

Measuring rhythm and tempo

• For prominence, compare the f0, duration and intensity of syllables (IMport vs imPORT)
• Rhythm measures compare how variable different intervals are: between stressed syllables, between all syllables, or between vowels and consonants
• Tempo is simpler: count the syllables and divide by the duration

Why F1 comes second

Landline telephones pass only about 350 to 3,400 Hz. This tends to raise F1 estimates, especially in close vowels with low F1. We come back to recording quality in session 4.

Which features identify a speaker best?

Experts’ answers, most often named first:

1. Voice quality
2. Dialect and accent
3. Vowel formants
4. Speech tempo, f0, rhythm, lexical and grammatical choices, fluency

Exercises

The book’s exercises for this chapter use a short recording of the sentence When the sunlight strikes raindrops in the air, they act as a prism and form a rainbow:

• Annotate sounds in Praat
• Measure F1, F2 and F3 at the midpoints of vowels
• Calculate the speaker’s articulation rate

Summary

Slides and code

These slides

mcanzi.de/ws27_fss/02_phonetic_analysis.html

Code and all sessions

github.com/mcanzi/ws27_fss